Estimate Human Pose For Multiple Person Using Pretrained Network

Load pretrained pose estimator model

detector = posenet.PoseEstimator;

Next, we will try to estimate keypoints with a test image. First, read a test image.

I = imread('visionteam1.jpg');

imshow(I);

Detect people in the loaded image.

[bboxes,scores] = detectPeopleACF(I);

Iout = insertObjectAnnotation(I,'rectangle',bboxes,scores,'LineWidth',3);

imshow(Iout)

title('Detected people and detection scores')

Extract and transform the detected objects to fit the input of the network.

[croppedImages, croppedBBoxes] = detector.normalizeBBoxes(I, bboxes);

Iout2 = insertObjectAnnotation(I,'rectangle',croppedBBoxes,scores,'LineWidth',3);

imshow(Iout2);

title('Resize the bounding boxes to be the same aspect ratio in the input of the network.');

figure, montage(croppedImages);

title('Each cropped image')

Estimate keypoints for each cropped image

heatmaps = detector.predict(croppedImages);

Iheatmaps = detector.visualizeHeatmaps(heatmaps, croppedImages);

montage(Iheatmaps);

title("Joint heatmaps")

keypoints = detector.heatmaps2Keypoints(heatmaps);

Iheatmaps = detector.visualizeKeyPoints(Iheatmaps,keypoints);

montage(Iheatmaps);

title('Extracted joints for each person');

Iout3 = detector.visualizeKeyPointsMultiple(I,keypoints,croppedBBoxes);

imshow(Iout3);

title('Estimated keypoints in the coordinates of the original image');